RESOLUTION- AND STIMULUS-AGNOSTIC SUPER-RESOLUTION OF ULTRA-HIGH-FIELD FUNCTIONAL MRI: APPLICATION TO VISUAL STUDIES.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Brain Mapping

Background:

• High-resolution functional MRI (fMRI) is crucial for understanding brain mesoscale organization.
• Increased spatial resolution in fMRI necessitates longer scan times due to low signal-to-noise ratios.
• Current methods often require sub-millimeter isotropic data for detailed analysis.

Purpose of the Study:

• To develop and validate a deep learning-based 3D super-resolution (SR) method for fMRI.
• To enable visualization of fine-scale neural organization using lower-resolution fMRI data.
• To reduce fMRI acquisition time while maintaining or improving spatial resolution.

Main Methods:

• A novel deep learning-based 3D super-resolution algorithm for fMRI data was developed.
• A resolution-agnostic image augmentation framework allows adaptation to various voxel sizes without retraining.
• The method was applied to localize motion-selective sites in early visual areas using 2-3mm isotropic fMRI data.

Main Results:

• The super-resolution fMRI successfully recovered high-frequency details of interdigitated motion-selective sites.
• The method demonstrated robustness and versatility, utilizing training data from different subjects and paradigms, including resting-state fMRI.
• Quantitative and qualitative analyses confirmed the enhancement of spatial resolution and potential for reduced scan times.

Conclusions:

• The developed deep learning SR method significantly enhances fMRI spatial resolution.
• This technique allows for the detection of fine-scale neural organization with lower-resolution, faster scans.
• The approach offers a promising avenue for reducing fMRI acquisition time and improving brain imaging efficiency.